GTASA-01: Multi-Actor Video Corpus with Perfect Spatiotemporal Annotations

GTASA-01 is the corpus of the GTASA paper GTASA: Ground Truth Annotations for Spatiotemporal Analysis, Evaluation and Training of Video Models (ACCV 2026), generated with the GEST-Engine described in the accompanying technical report.

The repository contains 1,008 procedurally generated multi-actor stories produced by the GEST-Engine, each accompanied by a Graph of Events in Space and Time (GEST) specification, an engine-rendered RGB video with dense spatiotemporal annotations, and — for comparison — videos produced by VEO 3.1 and WAN 2.2 from the same textual prompt.

Contents and paper subsets

The stories are stored in three generation batches:

Folder	Stories
segmentations_balanced_1/	398
segmentations_balanced_2/	306
segmentations_balanced_3/	304
Total	1,008

• GTASA (ACCV 2026) — the 938-scenario corpus reported in the paper, drawn from all three batches (373 / 286 / 279). The exact list of story folders is in gtasa_938_stories.txt.
• ICLR 2026 Tiny Paper — the 398-story sample released with GEST-Engine: Controllable Multi-Actor Video Synthesis with Perfect Spatiotemporal Annotations is segmentations_balanced_1/.
• The remaining 70 stories are not part of the GTASA paper set; 27 of them lack texts.json and the VEO / WAN reference videos.

Each batch folder also contains batch_report.md, batch_statistics.json and batch_summary.json describing that generation batch.

Corpus statistics

GTASA corpus (938 scenarios), as reported in the paper:

Metric	Value
Scenarios (Video–GEST–Text triplets)	938
Reference videos (VEO, WAN)	938 each
Total duration	14.25 hours (median 48.7 s per video)
Action clips	28.2K
Annotated frames	1.45M
Events, with exact frame mappings	29,579
Temporal relations	12,457
Pairwise spatial relations	892M
Actors per scenario	2–6 (mean 3.43)
Events per scenario	7–65 (mean 29.4)
Words per description	153 (mean)
Unique action types	37 (social, manipulation, locomotion, exercise)
Object types	15 (furniture, devices, consumables, equipment)
Environments	10 episodes, 4 categories

ICLR 2026 Tiny Paper sample (398 stories, segmentations_balanced_1/): 11,627 events, 4,603 temporal relations (43.5% before, 43.5% after, 13% same_time), 2–6 actors per story (mean 3.38), 10–65 events per story (mean 29.21).

Directory structure

gtasa_938_stories.txt                  # story folders in the GTASA (ACCV 2026) corpus
segmentations_balanced_{1,2,3}/
├── batch_report.md
├── batch_statistics.json
├── batch_summary.json
└── <story_folder>/ ...

Each story folder is named after its generation configuration (e.g., classroom_max2actors_max1regions_2action_chains_b9d2ff0d) and follows this layout:

<story_folder>/
├── texts.json                         # GPT-4o query + refined natural-language description
├── veo3-1.mp4                         # VEO 3.1 video from the refined description
├── wan2.2.mp4                         # WAN 2.2 video from the refined description
│
├── detailed_graph/
│   └── take1/
│       ├── detail_gest.json           # the input GEST specification
│       └── proto-graph.json           # normalized-ID GEST with populated timeframes
│
└── simulations/
    └── take1_sim1/
        ├── event_frame_mapping.json   # {event_id → [startFrame, endFrame]} alignments
        │
        ├── camera1/
        │   ├── raw.mp4                # engine-rendered RGB video
        │   ├── segmentation_frames.zip     # per-frame HLSL instance segmentation masks
        │   ├── segmentation_mapping.json   # texture hash → story-level entity ID
        │   └── spatial_relations.zip       # per-frame pairwise spatial relation graphs
        │
        ├── logs/
        │   ├── clientscript.log       # MTA client-side script log
        │   └── server.log             # MTA server-side script log
        │
        └── textual_description/
            ├── engine_generated.txt   # Logger running commentary during simulation
            └── prompt.txt             # proto-language with GPT-4o instructions

File descriptions

Top-level (per story)

• texts.json — Output of the two-stage text generation pipeline (proto-language + LLM refinement). Contains the GPT-4o query (the proto-language paragraph wrapped in the refinement instruction) and the refined natural-language description returned by GPT-4o (gpt-4o-2024-08-06). The refined description is what was used to prompt VEO 3.1 and WAN 2.2.
• veo3-1.mp4 — Video generated by VEO 3.1 using the refined description from texts.json as prompt.
• wan2.2.mp4 — Video generated by WAN 2.2 using the same refined description as prompt.

texts.json, veo3-1.mp4 and wan2.2.mp4 are present for 981 of the 1,008 stories.

detailed_graph/take1/

• detail_gest.json — The Graph of Events in Space and Time specification for this story, including actor and object Exists nodes, per-event action / entities / location / timeframe / properties, and the temporal / spatial / semantic / camera relation sections.
• proto-graph.json — Intermediate transformation of the GEST used by the text generation pipeline: entity identifiers are normalized to a canonical format (e.g., a0 → actor0; spawnable IDs become id:0.0-class:mobilephone), and each event's Timeframe field is populated with the exact [startFrame, endFrame] range from the event-frame mapping.

simulations/take1_sim1/

• event_frame_mapping.json — Exact frame-level alignment of each GEST event to its start/end frame in the rendered video ({eventId → [startFrame, endFrame]}), with FPS metadata.

simulations/take1_sim1/camera1/

• raw.mp4 — RGB video of the multi-actor simulation rendered by the engine.
• segmentation_frames.zip — Per-frame instance segmentation masks produced via an HLSL shader with FNV-1a texture hashing.
• segmentation_mapping.json — Mapping from texture hash values to story-level entity IDs, linking segmentation masks back to the GEST specification.
• spatial_relations.zip — Per-frame pairwise spatial relation graphs (one JSON per frame). For each entity, each frame records: 3D position and rotation; camera-relative distance, horizontal and vertical angles, coarse direction bucket (front/back/left/right/above/below/combinations), and in-FOV flag; object type and model ID. Entities are tagged with their story-level storyObjectId, linking back to the input GEST. The camera state (position, lookAt, FOV, roll) is also stored per frame.

simulations/take1_sim1/logs/

• clientscript.log — Client-side Multi Theft Auto script log.
• server.log — Server-side Multi Theft Auto script log.

simulations/take1_sim1/textual_description/

• engine_generated.txt — Running-commentary text produced by the engine's Logger during simulation, reporting actions as they execute.
• prompt.txt — The proto-language paragraph (ungrammatical verb forms like sitdowns, takeouts, assembled mechanically from the proto-graph) wrapped with the instruction prompt sent to GPT-4o.

Source code

• GEST-Engine (simulation system): github.com/ncudlenco/mta-sim
• Procedural GEST generator and batch production orchestrator: github.com/ncudlenco/multiagent_story_system

Both repositories are tagged at v1.0-iclr2026, the exact state used to generate this corpus.

License and intellectual property notice

This dataset is released under CC BY-NC 4.0 for non-commercial academic research purposes only.

The videos in this corpus contain frames rendered by Grand Theft Auto: San Andreas (Rockstar Games / Take-Two Interactive, 2004) via the Multi Theft Auto modification framework. All in-game assets (3D models, textures, animations, environments) remain the property of their respective owners. We do not claim ownership of any Rockstar Games / Take-Two Interactive intellectual property. Use of this dataset is governed by both the CC BY-NC 4.0 license and applicable copyright law regarding the underlying game content.

Citation

If you use this corpus, please cite the GTASA paper (accepted at ACCV 2026) and, as appropriate, its supplementary technical report on the GEST-Engine:

@article{cudlenco2026gtasa,
  title={GTASA: Ground Truth Annotations for Spatiotemporal Analysis, Evaluation and Training of Video Models},
  author={Cudlenco, Nicolae and Masala, Mihai and Leordeanu, Marius},
  journal={arXiv preprint arXiv:2604.10385},
  year={2026}
}

@article{cudlenco2026gest,
  title={The GEST-Engine: From Event Graphs to Synthetic Video. A Full Technical Report},
  author={Cudlenco, Nicolae and Masala, Mihai and Leordeanu, Marius},
  journal={arXiv preprint arXiv:2607.12231},
  year={2026}
}

Related work using this corpus:

• Authoring for Living Worlds: Tool-Constrained LLM Agents for Executable Multi-Actor Scenarios, accepted at the ECCV 2026 Workshop on Agents in Living Worlds.
• [Tiny Paper] GEST-Engine: Controllable Multi-Actor Video Synthesis with Perfect Spatiotemporal Annotations, ICLR 2026 Workshop on World Models.

@article{cudlenco2026authoring,
  title={Authoring for Living Worlds: Tool-Constrained LLM Agents for Executable Multi-Actor Scenarios},
  author={Cudlenco, Nicolae and Masala, Mihai and Leordeanu, Marius},
  journal={arXiv preprint arXiv:2604.10383},
  year={2026}
}

@inproceedings{cudlenco2026tiny,
  title={[Tiny Paper] {GEST}-Engine: Controllable Multi-Actor Video Synthesis with Perfect Spatiotemporal Annotations},
  author={Nicolae Cudlenco and Mihai Masala and Marius Leordeanu},
  booktitle={ICLR 2026 the 2nd Workshop on World Models: Understanding, Modelling and Scaling},
  year={2026},
  url={https://openreview.net/forum?id=uUofPYVMZH}
}

Contact

For questions or issues, please open an issue on the GEST-Engine repository or contact nicolae.cudlenco@gmail.com.